Here, a novel co-occurrence network inference method for microbiome data is presented. It uses a goodness of fit to proportionality measure and random permutation to calculate theoretical p-values. The method was compared with 8 other pre-existing inference methods using 10 simulated data tables each with 80 samples and 524 species and five ecological data tables with 50 samples and upto 1280 species. The proportionality method was one of the best methods for the ecological data tables with low sparsity: it had 0.907 precision, 0.999 specificity, and 0.259 sensitivity, while the second best method, LSA, had 0.586 precision, 1.000 specificity, 0.280 sensitivity. The method was successfully applied to a stroke case control microbiome. It detected genera known to be associated with stroke, and highlighted a potentially interesting interaction between Pseudomonas and Aquamonas. Although, it has a limitation as the input dataset has to be low in sparsity, the novel method is very precise and specific technique for network inference with moderate sensitivity, and can be used as a hypothesis generator for microbiome research.
Microbiome is a complex community of microorganisms (bacteria) with various types of interactions such as competition and parasitism. Research on microbiome and interactions within the community can provide an additional layer of information. A novel method to generate a network with significant interactions in a microbiome is presented here. The method uses proportionality measure and random permutation to calculate how significant the interaction is. It was compared with several other existing methods such as Pearson correlation. The result suggests that the proportionality method is very precise and specific and it outperforms the other methods for some datasets. Stroke case control microbiome datasets were used to show that the method can be used as a hypothesis generator as it detected microorganisms associated with stroke and also interesting interactions that could be related to stroke.